Source-linked AI summary
What Should Agents Say? Action-state Communication for Efficient Multi-Agent Systems
Chen Huang, Yuhao Wu, Wenxuan Zhang
TL;DR
MAS communication often leaves unconstrained outputs in shared history, increasing token and context costs. The paper evaluates five strategies across two topologies and proposes PACT, which converts outputs into compact action-state records. Across controlled settings and coding harnesses, PACT improves the performance–cost trade-off with fewer tokens while preserving or improving performance.
Problem
Unconstrained inter-agent outputs are repeatedly reprocessed in shared history, increasing token costs and potentially exhausting context.
Method
PACT projects each raw nonterminal output into a compact public record containing receiver-facing action, state, and result fields.
Results
PACT reduces token usage by 38.7% on average across baselines and model scales while preserving or improving task performance.
Takeaways & Limitations
PACT offers a training-free, harness-agnostic communication protocol that improves the performance–cost trade-off across controlled MAS settings and production coding harnesses.
Takeaways & Limitations
Benefits are not fully explored for short-interaction systems or collaboration forms such as open-ended debate, tool-heavy planning, and dynamically routed networks.
Abstract
from arXiv · showhide
Multi-agent systems (MAS) built on large language models are typically organized around roles, pipelines, and turn schedules, while the content that agents pass to one another is often left as unconstrained natural language. However, this free-form communication can rapidly inflate token usage, consume the shared context window, and ultimately affect both system performance and inference cost. We analyze five common inter-agent communication strategies across two MAS topologies, finding that no fixed strategy is universally optimal. Instead, effective inter-agent messages consistently preserve action-centered information needed by downstream agents. Building on this, we propose the PACT (Protocolized Action-state Communication and Transmission), which treats inter-agent communication as a public state-update problem and projects each raw agent output into a compact action-state record before it enters shared history. Across different MAS topologies, PACT consistently improves the performance-cost trade-off, achieving comparable or stronger task performance with substantially fewer tokens. The gains extend to production coding harnesses: PACT lifts OpenHands' resolve rate at -10% tokens-per-resolved, and is resolve-neutral on SWE-agent while halving input tokens. Our code is publicly available at https://github.com/iNLP-Lab/PACT.
1 Introduction
MAS collaboration improves difficult-task solving but creates substantial token and context costs when agents repeatedly exchange unconstrained outputs. The paper identifies message content as a central design lever and introduces PACT to preserve action-relevant information compactly.
- MAS coordinates multiple model calls for complex tasks, but repeated inter-agent outputs consume far more tokens than single-shot solving.
- Verbose reasoning traces are repeatedly reprocessed by downstream agents, compounding token costs and potentially exhausting context before task completion.
- Existing MAS research emphasizes roles, schedules, and tool allocation while leaving inter-agent messages largely unconstrained.
- Across five strategies and two MAS settings, no fixed communication strategy is uniformly optimal, making transmitted content an underexplored design dimension.
- PACT projects each nonterminal output into a compact public action-state message before adding it to shared history.
- 38.7% average token reduction across baselines and model scales preserves or improves task performance in controlled MAS experiments.
2 Related Work
Prior MAS work largely studies orchestration—who talks and when—while paying less attention to what agents communicate. PACT addresses this neglected message-content dimension by making action, grounded state, and reusable results public while keeping deliberation private.
- MAS research has explored role specialization, debate, conversation frameworks, and output aggregation, but commonly leaves message content free-form.
- Repeated context reuse makes token costs grow because downstream agents reprocess earlier outputs, especially forwarded deliberative reasoning.
- PACT shifts attention from which agents communicate to what content should enter shared history.
- The protocol exposes only action, grounded state, and reusable result while excluding private deliberation from the inter-agent channel.
- The paper extends this communication focus beyond research scaffolds to real-world software-engineering harnesses such as OpenHands and SWE-agent.
3 Diagnostic Analysis of Inter-Agent Communication
The diagnostic study compares five communication strategies across split-evidence interaction and sequential-pipeline settings at multiple model scales. It finds that useful messages are action-centered but must retain enough state and handoff information, motivating compact structured communication.
- Experimental design: The study holds other MAS factors fixed while comparing five communication strategies across two structurally different settings and multiple model scales.
- Setting A — Split-Evidence Interaction: Setting A splits evidence between two agents that must exchange task-relevant information over multiple turns.
- Setting B — Sequential pipeline: Setting B uses a four-agent Planner–Critic–Refiner–Solver pipeline in fixed order, with each stage building on prior intermediate artifacts.
- Communication strategies: The five strategies are Full Content, Concise Generation, Conclusion Only, Brief Summary, and Artifact Only.
- Diagnostic findings: No common strategy is uniformly satisfactory; useful communication is usually action-centered but must expose grounding state for reliable handoffs.
- Diagnostic findings: Full Content is consistently expensive and rarely best, with the sequential pipeline showing its largest token cost at every model scale.
- Diagnostic findings: Concise Generation works well for interaction but is weak in the role-specialized pipeline, while Brief Summary is unstable across scales.
- Diagnostic findings: Artifact Only often identifies the useful content, yet can require very large messages or extra turns because it omits state and closure cues.
4 PACT: Protocolized Action-state Communication and Transmission
PACT projects each agent’s raw output into a compact public action-state record before it enters shared history. It preserves receiver-facing action, grounded state, and reusable result while leaving private reasoning unconstrained.
- Protocol design: PACT makes handoffs explicit and grounded by separating sender intent, supporting evidence, and the result needed for continuation.The format prevents receivers from inferring intent or consuming unsupported conclusions.
- Communication boundary: Standard shared-history updates expose reasoning traces, explanations, repeated statements, and final artifacts to later agents.Because these elements are retained and reread, the raw-output update makes every part of the sender’s output public.
- Protocol design: PACT replaces raw inter-agent outputs with compact public action-state messages before appending them to shared history.A sender-side projection determines what information enters the inter-agent channel.
- Protocol design: Each PACT message contains ACTION, STATE, and RESULT fields for receiver-facing communication.ACTION specifies what was done or what should happen next; STATE supplies grounding evidence; RESULT carries the downstream artifact.
- Communication boundary: PACT excludes hidden reasoning traces and redundant restatements from public communication while preserving agents’ private computation.Agents may still reason, self-check, and explore alternatives; only the public inter-agent message is constrained.
5 Experiments
Experiments compare PACT with communication baselines across split-evidence and sequential-pipeline MAS settings. PACT shifts the performance–cost frontier by using compact, action-related handoffs, with especially large token savings against extended deliberation.
- Main results: PACT shifts the performance–cost frontier across the evaluated split-evidence and sequential-pipeline MAS settings.The experiments report performance and token cost across the two settings, with PACT achieving the most favorable trade-off.
- Split-evidence interaction: Full-history TextMAS uses more tokens than PACT and performs worse on both split-evidence benchmarks.The result indicates that retaining more conversation text does not necessarily create a more useful handoff.
- Sequential pipeline: 19%–23% of Multi-Agent Debate’s tokens are required by PACT in the sequential pipeline while achieving comparable or better performance on most benchmarks.The pipeline benefits from compact action-related artifacts that downstream agents can directly use.
- Scale trend: 21.1% lower average token usage and 4.2-point higher mean accuracy occur under PACT from Qwen3-8B to Qwen3-32B in the sequential pipeline.PACT token usage also decreases with model scale in the interaction setting.
- Ablation study: Removing both ACTION and STATE leaves only RESULT, reducing F1 to 64.3 and increasing token usage by 12.9%.The ablation identifies the complete action-state handoff as the source of explicit usage guidance, evidence grounding, and factual contribution.
6 PACT on Agentic Coding Harnesses
PACT transfers action-state communication to production coding harnesses through a lightweight proxy intervention, reducing redundant shared-history context while preserving or improving effectiveness. Across OpenHands and SWE-agent, it shifts the efficiency frontier toward lower communication cost.
- Setup: PACT is tested on OpenHands and SWE-agent over SWE-bench Verified using a proxy-level communication intervention.The proxy updates public messages between turns while leaving the underlying harness interfaces unchanged.
- Results: On OpenHands, resolve rate increases from 19.40% to 23.00% (+18 resolved instances), while tokens-per-resolved decreases by 10.3%.Average completion tokens per call also decrease by 5.3%.
- Results: On SWE-agent, input tokens decrease from 314.6 M to 156.0 M (−50.4%), while tokens-per-resolved drops by approximately 47%.The passage reports only a small resolve-rate change in this long-context loop.
- Interpretation: The action-state principle remains effective across external coding agents with different interaction histories and tool interfaces.PACT improves OpenHands and is approximately resolve-neutral on SWE-agent while reducing cost per solved instance.
7 Conclusion
The paper argues that inter-agent messages should preserve compact, action-centered public state rather than unconstrained intermediate content. PACT applies this principle across controlled MAS settings and production coding harnesses to improve the performance–cost trade-off.
- Conclusion: No common communication strategy is uniformly optimal across the two analyzed MAS settings, while useful inter-agent content is usually action-centered.The diagnostic analysis compares five common communication strategies across a split-evidence interaction and a sequential pipeline.
- Conclusion: PACT restricts shared history to compact action-state messages while excluding intermediate content without affecting the agent model’s capability.The protocol updates public message content rather than the agent’s private computation.
- Conclusion: Across controlled MAS settings and production coding harnesses, PACT reduces redundant context while preserving or improving task performance.The conclusion presents disciplining the shared inter-agent channel as an effective mechanism for reducing redundant context growth.
Limitations
PACT’s benefits are established in settings where shared conversational history is a major token-cost source, but its behavior in shorter or structurally different interactions remains underexplored.
- Scope boundary: PACT’s benefits are not fully explored for systems with short interactions or architectures that do not repeatedly expose prior agent outputs.The stated design target is MAS where shared conversational history is a major source of token cost.
- Scope boundary: The experiments cover two controlled MAS topologies and two agentic coding harnesses rather than all forms of multi-agent collaboration.Open-ended debate, tool-heavy planning, and dynamically routed agent networks are named as uncovered collaboration forms.
A Experimental Details
The experiments evaluate PACT and communication baselines across multiple model scales, benchmarks, MAS topologies, and external coding harnesses using specified decoding and interaction settings.
- Models: The evaluation uses Qwen3 models at 8B, 14B, and 32B scales.These scales cover mid-scale and large-scale reasoning models.
- Benchmarks: The split-evidence interaction uses HotpotQA and 2WikiMultiHopQA, while the sequential pipeline covers AIME2024, AIME2025, GPQA-Diamond, and OpenBookQA.The benchmarks span mathematical, scientific, and commonsense reasoning.
- Diagnostic analysis: The diagnostic analysis compares Full Content, Concise Generation, Conclusion Only, Brief Summary, and Artifact Only in both MAS settings.The split-evidence interaction uses a 5–5 context split, at most 8 turns, early exit, and max_new_tokens of 8,192.
- PACT experiments: PACT experiments reuse the four-agent Planner→Critic→Refiner→Solver pipeline and the two-agent split-evidence interaction.The split-evidence setting uses four alternating turns, max_new_tokens=4,096 per turn, and a 5–5 split.
- Baselines: The baselines differ only in inter-agent communication protocol while sharing decoding settings, token budgets, and dataset splits.They include Chain of Agents, TextMAS, and Multi-Agent Debate under the corresponding MAS configurations.
- Agentic coding harnesses: OpenHands and SWE-agent are evaluated on SWE-bench Verified with Qwen3-14B, applying PACT through a LiteLLM proxy hook without SDK or docker-image changes.The proxy rewrites prompts and parses responses in flight.
B Split-Evidence Interaction Case Study: How PACT Changes the Message
Figure 5 contrasts a full-deliberation message with PACT’s compact action-state message in a HotpotQA bridge-question interaction. PACT preserves the surfaced fact and explicit request while making the message approximately 10× shorter, keeping the same actionable content.
- PACT produces an approximately 10× shorter message while preserving the same actionable content.The compact message retains the surfaced fact and precise request rather than the full deliberation.
- Without PACT, the agent forwards its full <think> deliberation and hedged prose, requiring the partner to reread hundreds of tokens.
- PACT keeps the surfaced fact and precise request, so downstream turns stay accurate while shared history stops compounding.